OpenAI Astra's First Act: Ten Math Problems Stuck for a Decade
OpenAI's next major model has a name now, Astra, and its first public act isn't a chatbot demo. It's math. On August 1 OpenAI published ten advances in mathematics and theoretical computer science, each on a problem where the main result hadn't moved in at least ten years. An internal version of Astra generated the arguments, human mathematicians turned them into manuscripts, and everything was formalized in Lean, so the proofs check mechanically instead of resting on "trust the model". The haul includes new upper bounds for high-dimensional sphere packing down to the Cohn-Elkies threshold, a construction of non-sofic groups, polynomial-factor hardness of approximation for the closest vector problem (this one touches post-quantum crypto), plus results on Ehrhart's conjecture and multicolor Ramsey numbers, including a resolution of Erdos problem 183.
The number that matters most: OpenAI says the total model usage needed to find all ten solutions would cost about $2,000 at GPT-5.6 Sol API rates. Two thousand dollars for ten pieces of math the field left on the table for a decade. You can distrust frontier-lab benchmark claims all you want, but Lean certificates are the kind of receipt you can't fake, and the whole package is public: a 249-page collection of arguments, 62 pages of narrated discovery notes, and a repository of machine-checkable proofs. Formal peer review is still pending, and real humans did real work preparing the manuscripts, so "AI solved math alone" oversells it. But "AI found the key idea and humans wrote it down" is exactly how collaboration between strong mathematicians already works.
Context makes this bigger. Reports describe Astra as a model family built for multiple agents working one problem for hours or days, and Sam Altman has already been demoing it to policymakers. So this drop is a positioning move: before the model even launches, OpenAI wants "discovers new mathematics" to be its first impression, the same way Anthropic has been pushing autonomous research results from Claude. The frontier is quietly shifting from answering questions to producing new knowledge, and both labs know whoever makes that stick first owns the narrative.
Announcement: https://openai.com/index/ten-advances-in-mathematics/ and Simon Willison's notes: https://simonwillison.net/2026/Aug/1/ten-advances-in-mathematics/
← Back to all articles
The number that matters most: OpenAI says the total model usage needed to find all ten solutions would cost about $2,000 at GPT-5.6 Sol API rates. Two thousand dollars for ten pieces of math the field left on the table for a decade. You can distrust frontier-lab benchmark claims all you want, but Lean certificates are the kind of receipt you can't fake, and the whole package is public: a 249-page collection of arguments, 62 pages of narrated discovery notes, and a repository of machine-checkable proofs. Formal peer review is still pending, and real humans did real work preparing the manuscripts, so "AI solved math alone" oversells it. But "AI found the key idea and humans wrote it down" is exactly how collaboration between strong mathematicians already works.
Context makes this bigger. Reports describe Astra as a model family built for multiple agents working one problem for hours or days, and Sam Altman has already been demoing it to policymakers. So this drop is a positioning move: before the model even launches, OpenAI wants "discovers new mathematics" to be its first impression, the same way Anthropic has been pushing autonomous research results from Claude. The frontier is quietly shifting from answering questions to producing new knowledge, and both labs know whoever makes that stick first owns the narrative.
Announcement: https://openai.com/index/ten-advances-in-mathematics/ and Simon Willison's notes: https://simonwillison.net/2026/Aug/1/ten-advances-in-mathematics/
Comments